1 //===-- Host.cpp - Implement OS Host Concept --------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the operating system Host concept.
10 //
11 //===----------------------------------------------------------------------===//
12
13 #include "llvm/Support/Host.h"
14 #include "llvm/ADT/SmallSet.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringMap.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/StringSwitch.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/Config/llvm-config.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/FileSystem.h"
23 #include "llvm/Support/MemoryBuffer.h"
24 #include "llvm/Support/X86TargetParser.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include <assert.h>
27 #include <string.h>
28
29 // Include the platform-specific parts of this class.
30 #ifdef LLVM_ON_UNIX
31 #include "Unix/Host.inc"
32 #include <sched.h>
33 #endif
34 #ifdef _WIN32
35 #include "Windows/Host.inc"
36 #endif
37 #ifdef _MSC_VER
38 #include <intrin.h>
39 #endif
40 #if defined(__APPLE__) && (!defined(__x86_64__))
41 #include <mach/host_info.h>
42 #include <mach/mach.h>
43 #include <mach/mach_host.h>
44 #include <mach/machine.h>
45 #endif
46 #ifdef _AIX
47 #include <sys/systemcfg.h>
48 #endif
49
50 #define DEBUG_TYPE "host-detection"
51
52 //===----------------------------------------------------------------------===//
53 //
54 // Implementations of the CPU detection routines
55 //
56 //===----------------------------------------------------------------------===//
57
58 using namespace llvm;
59
60 static std::unique_ptr<llvm::MemoryBuffer>
getProcCpuinfoContent()61 LLVM_ATTRIBUTE_UNUSED getProcCpuinfoContent() {
62 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> Text =
63 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo");
64 if (std::error_code EC = Text.getError()) {
65 llvm::errs() << "Can't read "
66 << "/proc/cpuinfo: " << EC.message() << "\n";
67 return nullptr;
68 }
69 return std::move(*Text);
70 }
71
getHostCPUNameForPowerPC(StringRef ProcCpuinfoContent)72 StringRef sys::detail::getHostCPUNameForPowerPC(StringRef ProcCpuinfoContent) {
73 // Access to the Processor Version Register (PVR) on PowerPC is privileged,
74 // and so we must use an operating-system interface to determine the current
75 // processor type. On Linux, this is exposed through the /proc/cpuinfo file.
76 const char *generic = "generic";
77
78 // The cpu line is second (after the 'processor: 0' line), so if this
79 // buffer is too small then something has changed (or is wrong).
80 StringRef::const_iterator CPUInfoStart = ProcCpuinfoContent.begin();
81 StringRef::const_iterator CPUInfoEnd = ProcCpuinfoContent.end();
82
83 StringRef::const_iterator CIP = CPUInfoStart;
84
85 StringRef::const_iterator CPUStart = 0;
86 size_t CPULen = 0;
87
88 // We need to find the first line which starts with cpu, spaces, and a colon.
89 // After the colon, there may be some additional spaces and then the cpu type.
90 while (CIP < CPUInfoEnd && CPUStart == 0) {
91 if (CIP < CPUInfoEnd && *CIP == '\n')
92 ++CIP;
93
94 if (CIP < CPUInfoEnd && *CIP == 'c') {
95 ++CIP;
96 if (CIP < CPUInfoEnd && *CIP == 'p') {
97 ++CIP;
98 if (CIP < CPUInfoEnd && *CIP == 'u') {
99 ++CIP;
100 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
101 ++CIP;
102
103 if (CIP < CPUInfoEnd && *CIP == ':') {
104 ++CIP;
105 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
106 ++CIP;
107
108 if (CIP < CPUInfoEnd) {
109 CPUStart = CIP;
110 while (CIP < CPUInfoEnd && (*CIP != ' ' && *CIP != '\t' &&
111 *CIP != ',' && *CIP != '\n'))
112 ++CIP;
113 CPULen = CIP - CPUStart;
114 }
115 }
116 }
117 }
118 }
119
120 if (CPUStart == 0)
121 while (CIP < CPUInfoEnd && *CIP != '\n')
122 ++CIP;
123 }
124
125 if (CPUStart == 0)
126 return generic;
127
128 return StringSwitch<const char *>(StringRef(CPUStart, CPULen))
129 .Case("604e", "604e")
130 .Case("604", "604")
131 .Case("7400", "7400")
132 .Case("7410", "7400")
133 .Case("7447", "7400")
134 .Case("7455", "7450")
135 .Case("G4", "g4")
136 .Case("POWER4", "970")
137 .Case("PPC970FX", "970")
138 .Case("PPC970MP", "970")
139 .Case("G5", "g5")
140 .Case("POWER5", "g5")
141 .Case("A2", "a2")
142 .Case("POWER6", "pwr6")
143 .Case("POWER7", "pwr7")
144 .Case("POWER8", "pwr8")
145 .Case("POWER8E", "pwr8")
146 .Case("POWER8NVL", "pwr8")
147 .Case("POWER9", "pwr9")
148 .Case("POWER10", "pwr10")
149 // FIXME: If we get a simulator or machine with the capabilities of
150 // mcpu=future, we should revisit this and add the name reported by the
151 // simulator/machine.
152 .Default(generic);
153 }
154
getHostCPUNameForARM(StringRef ProcCpuinfoContent)155 StringRef sys::detail::getHostCPUNameForARM(StringRef ProcCpuinfoContent) {
156 // The cpuid register on arm is not accessible from user space. On Linux,
157 // it is exposed through the /proc/cpuinfo file.
158
159 // Read 32 lines from /proc/cpuinfo, which should contain the CPU part line
160 // in all cases.
161 SmallVector<StringRef, 32> Lines;
162 ProcCpuinfoContent.split(Lines, "\n");
163
164 // Look for the CPU implementer line.
165 StringRef Implementer;
166 StringRef Hardware;
167 StringRef Part;
168 for (unsigned I = 0, E = Lines.size(); I != E; ++I) {
169 if (Lines[I].startswith("CPU implementer"))
170 Implementer = Lines[I].substr(15).ltrim("\t :");
171 if (Lines[I].startswith("Hardware"))
172 Hardware = Lines[I].substr(8).ltrim("\t :");
173 if (Lines[I].startswith("CPU part"))
174 Part = Lines[I].substr(8).ltrim("\t :");
175 }
176
177 if (Implementer == "0x41") { // ARM Ltd.
178 // MSM8992/8994 may give cpu part for the core that the kernel is running on,
179 // which is undeterministic and wrong. Always return cortex-a53 for these SoC.
180 if (Hardware.endswith("MSM8994") || Hardware.endswith("MSM8996"))
181 return "cortex-a53";
182
183
184 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
185 // values correspond to the "Part number" in the CP15/c0 register. The
186 // contents are specified in the various processor manuals.
187 // This corresponds to the Main ID Register in Technical Reference Manuals.
188 // and is used in programs like sys-utils
189 return StringSwitch<const char *>(Part)
190 .Case("0x926", "arm926ej-s")
191 .Case("0xb02", "mpcore")
192 .Case("0xb36", "arm1136j-s")
193 .Case("0xb56", "arm1156t2-s")
194 .Case("0xb76", "arm1176jz-s")
195 .Case("0xc08", "cortex-a8")
196 .Case("0xc09", "cortex-a9")
197 .Case("0xc0f", "cortex-a15")
198 .Case("0xc20", "cortex-m0")
199 .Case("0xc23", "cortex-m3")
200 .Case("0xc24", "cortex-m4")
201 .Case("0xd22", "cortex-m55")
202 .Case("0xd02", "cortex-a34")
203 .Case("0xd04", "cortex-a35")
204 .Case("0xd03", "cortex-a53")
205 .Case("0xd07", "cortex-a57")
206 .Case("0xd08", "cortex-a72")
207 .Case("0xd09", "cortex-a73")
208 .Case("0xd0a", "cortex-a75")
209 .Case("0xd0b", "cortex-a76")
210 .Case("0xd0d", "cortex-a77")
211 .Case("0xd41", "cortex-a78")
212 .Case("0xd44", "cortex-x1")
213 .Case("0xd0c", "neoverse-n1")
214 .Case("0xd49", "neoverse-n2")
215 .Default("generic");
216 }
217
218 if (Implementer == "0x42" || Implementer == "0x43") { // Broadcom | Cavium.
219 return StringSwitch<const char *>(Part)
220 .Case("0x516", "thunderx2t99")
221 .Case("0x0516", "thunderx2t99")
222 .Case("0xaf", "thunderx2t99")
223 .Case("0x0af", "thunderx2t99")
224 .Case("0xa1", "thunderxt88")
225 .Case("0x0a1", "thunderxt88")
226 .Default("generic");
227 }
228
229 if (Implementer == "0x46") { // Fujitsu Ltd.
230 return StringSwitch<const char *>(Part)
231 .Case("0x001", "a64fx")
232 .Default("generic");
233 }
234
235 if (Implementer == "0x4e") { // NVIDIA Corporation
236 return StringSwitch<const char *>(Part)
237 .Case("0x004", "carmel")
238 .Default("generic");
239 }
240
241 if (Implementer == "0x48") // HiSilicon Technologies, Inc.
242 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
243 // values correspond to the "Part number" in the CP15/c0 register. The
244 // contents are specified in the various processor manuals.
245 return StringSwitch<const char *>(Part)
246 .Case("0xd01", "tsv110")
247 .Default("generic");
248
249 if (Implementer == "0x51") // Qualcomm Technologies, Inc.
250 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
251 // values correspond to the "Part number" in the CP15/c0 register. The
252 // contents are specified in the various processor manuals.
253 return StringSwitch<const char *>(Part)
254 .Case("0x06f", "krait") // APQ8064
255 .Case("0x201", "kryo")
256 .Case("0x205", "kryo")
257 .Case("0x211", "kryo")
258 .Case("0x800", "cortex-a73") // Kryo 2xx Gold
259 .Case("0x801", "cortex-a73") // Kryo 2xx Silver
260 .Case("0x802", "cortex-a75") // Kryo 3xx Gold
261 .Case("0x803", "cortex-a75") // Kryo 3xx Silver
262 .Case("0x804", "cortex-a76") // Kryo 4xx Gold
263 .Case("0x805", "cortex-a76") // Kryo 4xx/5xx Silver
264 .Case("0xc00", "falkor")
265 .Case("0xc01", "saphira")
266 .Default("generic");
267 if (Implementer == "0x53") { // Samsung Electronics Co., Ltd.
268 // The Exynos chips have a convoluted ID scheme that doesn't seem to follow
269 // any predictive pattern across variants and parts.
270 unsigned Variant = 0, Part = 0;
271
272 // Look for the CPU variant line, whose value is a 1 digit hexadecimal
273 // number, corresponding to the Variant bits in the CP15/C0 register.
274 for (auto I : Lines)
275 if (I.consume_front("CPU variant"))
276 I.ltrim("\t :").getAsInteger(0, Variant);
277
278 // Look for the CPU part line, whose value is a 3 digit hexadecimal
279 // number, corresponding to the PartNum bits in the CP15/C0 register.
280 for (auto I : Lines)
281 if (I.consume_front("CPU part"))
282 I.ltrim("\t :").getAsInteger(0, Part);
283
284 unsigned Exynos = (Variant << 12) | Part;
285 switch (Exynos) {
286 default:
287 // Default by falling through to Exynos M3.
288 LLVM_FALLTHROUGH;
289 case 0x1002:
290 return "exynos-m3";
291 case 0x1003:
292 return "exynos-m4";
293 }
294 }
295
296 return "generic";
297 }
298
getHostCPUNameForS390x(StringRef ProcCpuinfoContent)299 StringRef sys::detail::getHostCPUNameForS390x(StringRef ProcCpuinfoContent) {
300 // STIDP is a privileged operation, so use /proc/cpuinfo instead.
301
302 // The "processor 0:" line comes after a fair amount of other information,
303 // including a cache breakdown, but this should be plenty.
304 SmallVector<StringRef, 32> Lines;
305 ProcCpuinfoContent.split(Lines, "\n");
306
307 // Look for the CPU features.
308 SmallVector<StringRef, 32> CPUFeatures;
309 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
310 if (Lines[I].startswith("features")) {
311 size_t Pos = Lines[I].find(':');
312 if (Pos != StringRef::npos) {
313 Lines[I].drop_front(Pos + 1).split(CPUFeatures, ' ');
314 break;
315 }
316 }
317
318 // We need to check for the presence of vector support independently of
319 // the machine type, since we may only use the vector register set when
320 // supported by the kernel (and hypervisor).
321 bool HaveVectorSupport = false;
322 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) {
323 if (CPUFeatures[I] == "vx")
324 HaveVectorSupport = true;
325 }
326
327 // Now check the processor machine type.
328 for (unsigned I = 0, E = Lines.size(); I != E; ++I) {
329 if (Lines[I].startswith("processor ")) {
330 size_t Pos = Lines[I].find("machine = ");
331 if (Pos != StringRef::npos) {
332 Pos += sizeof("machine = ") - 1;
333 unsigned int Id;
334 if (!Lines[I].drop_front(Pos).getAsInteger(10, Id)) {
335 if (Id >= 8561 && HaveVectorSupport)
336 return "z15";
337 if (Id >= 3906 && HaveVectorSupport)
338 return "z14";
339 if (Id >= 2964 && HaveVectorSupport)
340 return "z13";
341 if (Id >= 2827)
342 return "zEC12";
343 if (Id >= 2817)
344 return "z196";
345 }
346 }
347 break;
348 }
349 }
350
351 return "generic";
352 }
353
getHostCPUNameForBPF()354 StringRef sys::detail::getHostCPUNameForBPF() {
355 #if !defined(__linux__) || !defined(__x86_64__)
356 return "generic";
357 #else
358 uint8_t v3_insns[40] __attribute__ ((aligned (8))) =
359 /* BPF_MOV64_IMM(BPF_REG_0, 0) */
360 { 0xb7, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
361 /* BPF_MOV64_IMM(BPF_REG_2, 1) */
362 0xb7, 0x2, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
363 /* BPF_JMP32_REG(BPF_JLT, BPF_REG_0, BPF_REG_2, 1) */
364 0xae, 0x20, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0,
365 /* BPF_MOV64_IMM(BPF_REG_0, 1) */
366 0xb7, 0x0, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
367 /* BPF_EXIT_INSN() */
368 0x95, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 };
369
370 uint8_t v2_insns[40] __attribute__ ((aligned (8))) =
371 /* BPF_MOV64_IMM(BPF_REG_0, 0) */
372 { 0xb7, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
373 /* BPF_MOV64_IMM(BPF_REG_2, 1) */
374 0xb7, 0x2, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
375 /* BPF_JMP_REG(BPF_JLT, BPF_REG_0, BPF_REG_2, 1) */
376 0xad, 0x20, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0,
377 /* BPF_MOV64_IMM(BPF_REG_0, 1) */
378 0xb7, 0x0, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
379 /* BPF_EXIT_INSN() */
380 0x95, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 };
381
382 struct bpf_prog_load_attr {
383 uint32_t prog_type;
384 uint32_t insn_cnt;
385 uint64_t insns;
386 uint64_t license;
387 uint32_t log_level;
388 uint32_t log_size;
389 uint64_t log_buf;
390 uint32_t kern_version;
391 uint32_t prog_flags;
392 } attr = {};
393 attr.prog_type = 1; /* BPF_PROG_TYPE_SOCKET_FILTER */
394 attr.insn_cnt = 5;
395 attr.insns = (uint64_t)v3_insns;
396 attr.license = (uint64_t)"DUMMY";
397
398 int fd = syscall(321 /* __NR_bpf */, 5 /* BPF_PROG_LOAD */, &attr,
399 sizeof(attr));
400 if (fd >= 0) {
401 close(fd);
402 return "v3";
403 }
404
405 /* Clear the whole attr in case its content changed by syscall. */
406 memset(&attr, 0, sizeof(attr));
407 attr.prog_type = 1; /* BPF_PROG_TYPE_SOCKET_FILTER */
408 attr.insn_cnt = 5;
409 attr.insns = (uint64_t)v2_insns;
410 attr.license = (uint64_t)"DUMMY";
411 fd = syscall(321 /* __NR_bpf */, 5 /* BPF_PROG_LOAD */, &attr, sizeof(attr));
412 if (fd >= 0) {
413 close(fd);
414 return "v2";
415 }
416 return "v1";
417 #endif
418 }
419
420 #if defined(__i386__) || defined(_M_IX86) || \
421 defined(__x86_64__) || defined(_M_X64)
422
423 // The check below for i386 was copied from clang's cpuid.h (__get_cpuid_max).
424 // Check motivated by bug reports for OpenSSL crashing on CPUs without CPUID
425 // support. Consequently, for i386, the presence of CPUID is checked first
426 // via the corresponding eflags bit.
427 // Removal of cpuid.h header motivated by PR30384
428 // Header cpuid.h and method __get_cpuid_max are not used in llvm, clang, openmp
429 // or test-suite, but are used in external projects e.g. libstdcxx
isCpuIdSupported()430 static bool isCpuIdSupported() {
431 #if defined(__GNUC__) || defined(__clang__)
432 #if defined(__i386__)
433 int __cpuid_supported;
434 __asm__(" pushfl\n"
435 " popl %%eax\n"
436 " movl %%eax,%%ecx\n"
437 " xorl $0x00200000,%%eax\n"
438 " pushl %%eax\n"
439 " popfl\n"
440 " pushfl\n"
441 " popl %%eax\n"
442 " movl $0,%0\n"
443 " cmpl %%eax,%%ecx\n"
444 " je 1f\n"
445 " movl $1,%0\n"
446 "1:"
447 : "=r"(__cpuid_supported)
448 :
449 : "eax", "ecx");
450 if (!__cpuid_supported)
451 return false;
452 #endif
453 return true;
454 #endif
455 return true;
456 }
457
458 /// getX86CpuIDAndInfo - Execute the specified cpuid and return the 4 values in
459 /// the specified arguments. If we can't run cpuid on the host, return true.
getX86CpuIDAndInfo(unsigned value,unsigned * rEAX,unsigned * rEBX,unsigned * rECX,unsigned * rEDX)460 static bool getX86CpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX,
461 unsigned *rECX, unsigned *rEDX) {
462 #if defined(__GNUC__) || defined(__clang__)
463 #if defined(__x86_64__)
464 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually.
465 // FIXME: should we save this for Clang?
466 __asm__("movq\t%%rbx, %%rsi\n\t"
467 "cpuid\n\t"
468 "xchgq\t%%rbx, %%rsi\n\t"
469 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX)
470 : "a"(value));
471 return false;
472 #elif defined(__i386__)
473 __asm__("movl\t%%ebx, %%esi\n\t"
474 "cpuid\n\t"
475 "xchgl\t%%ebx, %%esi\n\t"
476 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX)
477 : "a"(value));
478 return false;
479 #else
480 return true;
481 #endif
482 #elif defined(_MSC_VER)
483 // The MSVC intrinsic is portable across x86 and x64.
484 int registers[4];
485 __cpuid(registers, value);
486 *rEAX = registers[0];
487 *rEBX = registers[1];
488 *rECX = registers[2];
489 *rEDX = registers[3];
490 return false;
491 #else
492 return true;
493 #endif
494 }
495
496 namespace llvm {
497 namespace sys {
498 namespace detail {
499 namespace x86 {
500
getVendorSignature(unsigned * MaxLeaf)501 VendorSignatures getVendorSignature(unsigned *MaxLeaf) {
502 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
503 if (MaxLeaf == nullptr)
504 MaxLeaf = &EAX;
505 else
506 *MaxLeaf = 0;
507
508 if (!isCpuIdSupported())
509 return VendorSignatures::UNKNOWN;
510
511 if (getX86CpuIDAndInfo(0, MaxLeaf, &EBX, &ECX, &EDX) || *MaxLeaf < 1)
512 return VendorSignatures::UNKNOWN;
513
514 // "Genu ineI ntel"
515 if (EBX == 0x756e6547 && EDX == 0x49656e69 && ECX == 0x6c65746e)
516 return VendorSignatures::GENUINE_INTEL;
517
518 // "Auth enti cAMD"
519 if (EBX == 0x68747541 && EDX == 0x69746e65 && ECX == 0x444d4163)
520 return VendorSignatures::AUTHENTIC_AMD;
521
522 return VendorSignatures::UNKNOWN;
523 }
524
525 } // namespace x86
526 } // namespace detail
527 } // namespace sys
528 } // namespace llvm
529
530 using namespace llvm::sys::detail::x86;
531
532 /// getX86CpuIDAndInfoEx - Execute the specified cpuid with subleaf and return
533 /// the 4 values in the specified arguments. If we can't run cpuid on the host,
534 /// return true.
getX86CpuIDAndInfoEx(unsigned value,unsigned subleaf,unsigned * rEAX,unsigned * rEBX,unsigned * rECX,unsigned * rEDX)535 static bool getX86CpuIDAndInfoEx(unsigned value, unsigned subleaf,
536 unsigned *rEAX, unsigned *rEBX, unsigned *rECX,
537 unsigned *rEDX) {
538 #if defined(__GNUC__) || defined(__clang__)
539 #if defined(__x86_64__)
540 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually.
541 // FIXME: should we save this for Clang?
542 __asm__("movq\t%%rbx, %%rsi\n\t"
543 "cpuid\n\t"
544 "xchgq\t%%rbx, %%rsi\n\t"
545 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX)
546 : "a"(value), "c"(subleaf));
547 return false;
548 #elif defined(__i386__)
549 __asm__("movl\t%%ebx, %%esi\n\t"
550 "cpuid\n\t"
551 "xchgl\t%%ebx, %%esi\n\t"
552 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX)
553 : "a"(value), "c"(subleaf));
554 return false;
555 #else
556 return true;
557 #endif
558 #elif defined(_MSC_VER)
559 int registers[4];
560 __cpuidex(registers, value, subleaf);
561 *rEAX = registers[0];
562 *rEBX = registers[1];
563 *rECX = registers[2];
564 *rEDX = registers[3];
565 return false;
566 #else
567 return true;
568 #endif
569 }
570
571 // Read control register 0 (XCR0). Used to detect features such as AVX.
getX86XCR0(unsigned * rEAX,unsigned * rEDX)572 static bool getX86XCR0(unsigned *rEAX, unsigned *rEDX) {
573 #if defined(__GNUC__) || defined(__clang__)
574 // Check xgetbv; this uses a .byte sequence instead of the instruction
575 // directly because older assemblers do not include support for xgetbv and
576 // there is no easy way to conditionally compile based on the assembler used.
577 __asm__(".byte 0x0f, 0x01, 0xd0" : "=a"(*rEAX), "=d"(*rEDX) : "c"(0));
578 return false;
579 #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
580 unsigned long long Result = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
581 *rEAX = Result;
582 *rEDX = Result >> 32;
583 return false;
584 #else
585 return true;
586 #endif
587 }
588
detectX86FamilyModel(unsigned EAX,unsigned * Family,unsigned * Model)589 static void detectX86FamilyModel(unsigned EAX, unsigned *Family,
590 unsigned *Model) {
591 *Family = (EAX >> 8) & 0xf; // Bits 8 - 11
592 *Model = (EAX >> 4) & 0xf; // Bits 4 - 7
593 if (*Family == 6 || *Family == 0xf) {
594 if (*Family == 0xf)
595 // Examine extended family ID if family ID is F.
596 *Family += (EAX >> 20) & 0xff; // Bits 20 - 27
597 // Examine extended model ID if family ID is 6 or F.
598 *Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19
599 }
600 }
601
602 static StringRef
getIntelProcessorTypeAndSubtype(unsigned Family,unsigned Model,const unsigned * Features,unsigned * Type,unsigned * Subtype)603 getIntelProcessorTypeAndSubtype(unsigned Family, unsigned Model,
604 const unsigned *Features,
605 unsigned *Type, unsigned *Subtype) {
606 auto testFeature = [&](unsigned F) {
607 return (Features[F / 32] & (1U << (F % 32))) != 0;
608 };
609
610 StringRef CPU;
611
612 switch (Family) {
613 case 3:
614 CPU = "i386";
615 break;
616 case 4:
617 CPU = "i486";
618 break;
619 case 5:
620 if (testFeature(X86::FEATURE_MMX)) {
621 CPU = "pentium-mmx";
622 break;
623 }
624 CPU = "pentium";
625 break;
626 case 6:
627 switch (Model) {
628 case 0x0f: // Intel Core 2 Duo processor, Intel Core 2 Duo mobile
629 // processor, Intel Core 2 Quad processor, Intel Core 2 Quad
630 // mobile processor, Intel Core 2 Extreme processor, Intel
631 // Pentium Dual-Core processor, Intel Xeon processor, model
632 // 0Fh. All processors are manufactured using the 65 nm process.
633 case 0x16: // Intel Celeron processor model 16h. All processors are
634 // manufactured using the 65 nm process
635 CPU = "core2";
636 *Type = X86::INTEL_CORE2;
637 break;
638 case 0x17: // Intel Core 2 Extreme processor, Intel Xeon processor, model
639 // 17h. All processors are manufactured using the 45 nm process.
640 //
641 // 45nm: Penryn , Wolfdale, Yorkfield (XE)
642 case 0x1d: // Intel Xeon processor MP. All processors are manufactured using
643 // the 45 nm process.
644 CPU = "penryn";
645 *Type = X86::INTEL_CORE2;
646 break;
647 case 0x1a: // Intel Core i7 processor and Intel Xeon processor. All
648 // processors are manufactured using the 45 nm process.
649 case 0x1e: // Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz.
650 // As found in a Summer 2010 model iMac.
651 case 0x1f:
652 case 0x2e: // Nehalem EX
653 CPU = "nehalem";
654 *Type = X86::INTEL_COREI7;
655 *Subtype = X86::INTEL_COREI7_NEHALEM;
656 break;
657 case 0x25: // Intel Core i7, laptop version.
658 case 0x2c: // Intel Core i7 processor and Intel Xeon processor. All
659 // processors are manufactured using the 32 nm process.
660 case 0x2f: // Westmere EX
661 CPU = "westmere";
662 *Type = X86::INTEL_COREI7;
663 *Subtype = X86::INTEL_COREI7_WESTMERE;
664 break;
665 case 0x2a: // Intel Core i7 processor. All processors are manufactured
666 // using the 32 nm process.
667 case 0x2d:
668 CPU = "sandybridge";
669 *Type = X86::INTEL_COREI7;
670 *Subtype = X86::INTEL_COREI7_SANDYBRIDGE;
671 break;
672 case 0x3a:
673 case 0x3e: // Ivy Bridge EP
674 CPU = "ivybridge";
675 *Type = X86::INTEL_COREI7;
676 *Subtype = X86::INTEL_COREI7_IVYBRIDGE;
677 break;
678
679 // Haswell:
680 case 0x3c:
681 case 0x3f:
682 case 0x45:
683 case 0x46:
684 CPU = "haswell";
685 *Type = X86::INTEL_COREI7;
686 *Subtype = X86::INTEL_COREI7_HASWELL;
687 break;
688
689 // Broadwell:
690 case 0x3d:
691 case 0x47:
692 case 0x4f:
693 case 0x56:
694 CPU = "broadwell";
695 *Type = X86::INTEL_COREI7;
696 *Subtype = X86::INTEL_COREI7_BROADWELL;
697 break;
698
699 // Skylake:
700 case 0x4e: // Skylake mobile
701 case 0x5e: // Skylake desktop
702 case 0x8e: // Kaby Lake mobile
703 case 0x9e: // Kaby Lake desktop
704 case 0xa5: // Comet Lake-H/S
705 case 0xa6: // Comet Lake-U
706 CPU = "skylake";
707 *Type = X86::INTEL_COREI7;
708 *Subtype = X86::INTEL_COREI7_SKYLAKE;
709 break;
710
711 // Rocketlake:
712 case 0xa7:
713 CPU = "rocketlake";
714 *Type = X86::INTEL_COREI7;
715 *Subtype = X86::INTEL_COREI7_ROCKETLAKE;
716 break;
717
718 // Skylake Xeon:
719 case 0x55:
720 *Type = X86::INTEL_COREI7;
721 if (testFeature(X86::FEATURE_AVX512BF16)) {
722 CPU = "cooperlake";
723 *Subtype = X86::INTEL_COREI7_COOPERLAKE;
724 } else if (testFeature(X86::FEATURE_AVX512VNNI)) {
725 CPU = "cascadelake";
726 *Subtype = X86::INTEL_COREI7_CASCADELAKE;
727 } else {
728 CPU = "skylake-avx512";
729 *Subtype = X86::INTEL_COREI7_SKYLAKE_AVX512;
730 }
731 break;
732
733 // Cannonlake:
734 case 0x66:
735 CPU = "cannonlake";
736 *Type = X86::INTEL_COREI7;
737 *Subtype = X86::INTEL_COREI7_CANNONLAKE;
738 break;
739
740 // Icelake:
741 case 0x7d:
742 case 0x7e:
743 CPU = "icelake-client";
744 *Type = X86::INTEL_COREI7;
745 *Subtype = X86::INTEL_COREI7_ICELAKE_CLIENT;
746 break;
747
748 // Icelake Xeon:
749 case 0x6a:
750 case 0x6c:
751 CPU = "icelake-server";
752 *Type = X86::INTEL_COREI7;
753 *Subtype = X86::INTEL_COREI7_ICELAKE_SERVER;
754 break;
755
756 // Sapphire Rapids:
757 case 0x8f:
758 CPU = "sapphirerapids";
759 *Type = X86::INTEL_COREI7;
760 *Subtype = X86::INTEL_COREI7_SAPPHIRERAPIDS;
761 break;
762
763 case 0x1c: // Most 45 nm Intel Atom processors
764 case 0x26: // 45 nm Atom Lincroft
765 case 0x27: // 32 nm Atom Medfield
766 case 0x35: // 32 nm Atom Midview
767 case 0x36: // 32 nm Atom Midview
768 CPU = "bonnell";
769 *Type = X86::INTEL_BONNELL;
770 break;
771
772 // Atom Silvermont codes from the Intel software optimization guide.
773 case 0x37:
774 case 0x4a:
775 case 0x4d:
776 case 0x5a:
777 case 0x5d:
778 case 0x4c: // really airmont
779 CPU = "silvermont";
780 *Type = X86::INTEL_SILVERMONT;
781 break;
782 // Goldmont:
783 case 0x5c: // Apollo Lake
784 case 0x5f: // Denverton
785 CPU = "goldmont";
786 *Type = X86::INTEL_GOLDMONT;
787 break;
788 case 0x7a:
789 CPU = "goldmont-plus";
790 *Type = X86::INTEL_GOLDMONT_PLUS;
791 break;
792 case 0x86:
793 CPU = "tremont";
794 *Type = X86::INTEL_TREMONT;
795 break;
796
797 // Xeon Phi (Knights Landing + Knights Mill):
798 case 0x57:
799 CPU = "knl";
800 *Type = X86::INTEL_KNL;
801 break;
802 case 0x85:
803 CPU = "knm";
804 *Type = X86::INTEL_KNM;
805 break;
806
807 default: // Unknown family 6 CPU, try to guess.
808 // Don't both with Type/Subtype here, they aren't used by the caller.
809 // They're used above to keep the code in sync with compiler-rt.
810 // TODO detect tigerlake host from model
811 if (testFeature(X86::FEATURE_AVX512VP2INTERSECT)) {
812 CPU = "tigerlake";
813 } else if (testFeature(X86::FEATURE_AVX512VBMI2)) {
814 CPU = "icelake-client";
815 } else if (testFeature(X86::FEATURE_AVX512VBMI)) {
816 CPU = "cannonlake";
817 } else if (testFeature(X86::FEATURE_AVX512BF16)) {
818 CPU = "cooperlake";
819 } else if (testFeature(X86::FEATURE_AVX512VNNI)) {
820 CPU = "cascadelake";
821 } else if (testFeature(X86::FEATURE_AVX512VL)) {
822 CPU = "skylake-avx512";
823 } else if (testFeature(X86::FEATURE_AVX512ER)) {
824 CPU = "knl";
825 } else if (testFeature(X86::FEATURE_CLFLUSHOPT)) {
826 if (testFeature(X86::FEATURE_SHA))
827 CPU = "goldmont";
828 else
829 CPU = "skylake";
830 } else if (testFeature(X86::FEATURE_ADX)) {
831 CPU = "broadwell";
832 } else if (testFeature(X86::FEATURE_AVX2)) {
833 CPU = "haswell";
834 } else if (testFeature(X86::FEATURE_AVX)) {
835 CPU = "sandybridge";
836 } else if (testFeature(X86::FEATURE_SSE4_2)) {
837 if (testFeature(X86::FEATURE_MOVBE))
838 CPU = "silvermont";
839 else
840 CPU = "nehalem";
841 } else if (testFeature(X86::FEATURE_SSE4_1)) {
842 CPU = "penryn";
843 } else if (testFeature(X86::FEATURE_SSSE3)) {
844 if (testFeature(X86::FEATURE_MOVBE))
845 CPU = "bonnell";
846 else
847 CPU = "core2";
848 } else if (testFeature(X86::FEATURE_64BIT)) {
849 CPU = "core2";
850 } else if (testFeature(X86::FEATURE_SSE3)) {
851 CPU = "yonah";
852 } else if (testFeature(X86::FEATURE_SSE2)) {
853 CPU = "pentium-m";
854 } else if (testFeature(X86::FEATURE_SSE)) {
855 CPU = "pentium3";
856 } else if (testFeature(X86::FEATURE_MMX)) {
857 CPU = "pentium2";
858 } else {
859 CPU = "pentiumpro";
860 }
861 break;
862 }
863 break;
864 case 15: {
865 if (testFeature(X86::FEATURE_64BIT)) {
866 CPU = "nocona";
867 break;
868 }
869 if (testFeature(X86::FEATURE_SSE3)) {
870 CPU = "prescott";
871 break;
872 }
873 CPU = "pentium4";
874 break;
875 }
876 default:
877 break; // Unknown.
878 }
879
880 return CPU;
881 }
882
883 static StringRef
getAMDProcessorTypeAndSubtype(unsigned Family,unsigned Model,const unsigned * Features,unsigned * Type,unsigned * Subtype)884 getAMDProcessorTypeAndSubtype(unsigned Family, unsigned Model,
885 const unsigned *Features,
886 unsigned *Type, unsigned *Subtype) {
887 auto testFeature = [&](unsigned F) {
888 return (Features[F / 32] & (1U << (F % 32))) != 0;
889 };
890
891 StringRef CPU;
892
893 switch (Family) {
894 case 4:
895 CPU = "i486";
896 break;
897 case 5:
898 CPU = "pentium";
899 switch (Model) {
900 case 6:
901 case 7:
902 CPU = "k6";
903 break;
904 case 8:
905 CPU = "k6-2";
906 break;
907 case 9:
908 case 13:
909 CPU = "k6-3";
910 break;
911 case 10:
912 CPU = "geode";
913 break;
914 }
915 break;
916 case 6:
917 if (testFeature(X86::FEATURE_SSE)) {
918 CPU = "athlon-xp";
919 break;
920 }
921 CPU = "athlon";
922 break;
923 case 15:
924 if (testFeature(X86::FEATURE_SSE3)) {
925 CPU = "k8-sse3";
926 break;
927 }
928 CPU = "k8";
929 break;
930 case 16:
931 CPU = "amdfam10";
932 *Type = X86::AMDFAM10H; // "amdfam10"
933 switch (Model) {
934 case 2:
935 *Subtype = X86::AMDFAM10H_BARCELONA;
936 break;
937 case 4:
938 *Subtype = X86::AMDFAM10H_SHANGHAI;
939 break;
940 case 8:
941 *Subtype = X86::AMDFAM10H_ISTANBUL;
942 break;
943 }
944 break;
945 case 20:
946 CPU = "btver1";
947 *Type = X86::AMD_BTVER1;
948 break;
949 case 21:
950 CPU = "bdver1";
951 *Type = X86::AMDFAM15H;
952 if (Model >= 0x60 && Model <= 0x7f) {
953 CPU = "bdver4";
954 *Subtype = X86::AMDFAM15H_BDVER4;
955 break; // 60h-7Fh: Excavator
956 }
957 if (Model >= 0x30 && Model <= 0x3f) {
958 CPU = "bdver3";
959 *Subtype = X86::AMDFAM15H_BDVER3;
960 break; // 30h-3Fh: Steamroller
961 }
962 if ((Model >= 0x10 && Model <= 0x1f) || Model == 0x02) {
963 CPU = "bdver2";
964 *Subtype = X86::AMDFAM15H_BDVER2;
965 break; // 02h, 10h-1Fh: Piledriver
966 }
967 if (Model <= 0x0f) {
968 *Subtype = X86::AMDFAM15H_BDVER1;
969 break; // 00h-0Fh: Bulldozer
970 }
971 break;
972 case 22:
973 CPU = "btver2";
974 *Type = X86::AMD_BTVER2;
975 break;
976 case 23:
977 CPU = "znver1";
978 *Type = X86::AMDFAM17H;
979 if ((Model >= 0x30 && Model <= 0x3f) || Model == 0x71) {
980 CPU = "znver2";
981 *Subtype = X86::AMDFAM17H_ZNVER2;
982 break; // 30h-3fh, 71h: Zen2
983 }
984 if (Model <= 0x0f) {
985 *Subtype = X86::AMDFAM17H_ZNVER1;
986 break; // 00h-0Fh: Zen1
987 }
988 break;
989 case 25:
990 CPU = "znver3";
991 *Type = X86::AMDFAM19H;
992 if (Model <= 0x0f) {
993 *Subtype = X86::AMDFAM19H_ZNVER3;
994 break; // 00h-0Fh: Zen3
995 }
996 break;
997 default:
998 break; // Unknown AMD CPU.
999 }
1000
1001 return CPU;
1002 }
1003
getAvailableFeatures(unsigned ECX,unsigned EDX,unsigned MaxLeaf,unsigned * Features)1004 static void getAvailableFeatures(unsigned ECX, unsigned EDX, unsigned MaxLeaf,
1005 unsigned *Features) {
1006 unsigned EAX, EBX;
1007
1008 auto setFeature = [&](unsigned F) {
1009 Features[F / 32] |= 1U << (F % 32);
1010 };
1011
1012 if ((EDX >> 15) & 1)
1013 setFeature(X86::FEATURE_CMOV);
1014 if ((EDX >> 23) & 1)
1015 setFeature(X86::FEATURE_MMX);
1016 if ((EDX >> 25) & 1)
1017 setFeature(X86::FEATURE_SSE);
1018 if ((EDX >> 26) & 1)
1019 setFeature(X86::FEATURE_SSE2);
1020
1021 if ((ECX >> 0) & 1)
1022 setFeature(X86::FEATURE_SSE3);
1023 if ((ECX >> 1) & 1)
1024 setFeature(X86::FEATURE_PCLMUL);
1025 if ((ECX >> 9) & 1)
1026 setFeature(X86::FEATURE_SSSE3);
1027 if ((ECX >> 12) & 1)
1028 setFeature(X86::FEATURE_FMA);
1029 if ((ECX >> 19) & 1)
1030 setFeature(X86::FEATURE_SSE4_1);
1031 if ((ECX >> 20) & 1)
1032 setFeature(X86::FEATURE_SSE4_2);
1033 if ((ECX >> 23) & 1)
1034 setFeature(X86::FEATURE_POPCNT);
1035 if ((ECX >> 25) & 1)
1036 setFeature(X86::FEATURE_AES);
1037
1038 if ((ECX >> 22) & 1)
1039 setFeature(X86::FEATURE_MOVBE);
1040
1041 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
1042 // indicates that the AVX registers will be saved and restored on context
1043 // switch, then we have full AVX support.
1044 const unsigned AVXBits = (1 << 27) | (1 << 28);
1045 bool HasAVX = ((ECX & AVXBits) == AVXBits) && !getX86XCR0(&EAX, &EDX) &&
1046 ((EAX & 0x6) == 0x6);
1047 #if defined(__APPLE__)
1048 // Darwin lazily saves the AVX512 context on first use: trust that the OS will
1049 // save the AVX512 context if we use AVX512 instructions, even the bit is not
1050 // set right now.
1051 bool HasAVX512Save = true;
1052 #else
1053 // AVX512 requires additional context to be saved by the OS.
1054 bool HasAVX512Save = HasAVX && ((EAX & 0xe0) == 0xe0);
1055 #endif
1056
1057 if (HasAVX)
1058 setFeature(X86::FEATURE_AVX);
1059
1060 bool HasLeaf7 =
1061 MaxLeaf >= 0x7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX);
1062
1063 if (HasLeaf7 && ((EBX >> 3) & 1))
1064 setFeature(X86::FEATURE_BMI);
1065 if (HasLeaf7 && ((EBX >> 5) & 1) && HasAVX)
1066 setFeature(X86::FEATURE_AVX2);
1067 if (HasLeaf7 && ((EBX >> 8) & 1))
1068 setFeature(X86::FEATURE_BMI2);
1069 if (HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save)
1070 setFeature(X86::FEATURE_AVX512F);
1071 if (HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save)
1072 setFeature(X86::FEATURE_AVX512DQ);
1073 if (HasLeaf7 && ((EBX >> 19) & 1))
1074 setFeature(X86::FEATURE_ADX);
1075 if (HasLeaf7 && ((EBX >> 21) & 1) && HasAVX512Save)
1076 setFeature(X86::FEATURE_AVX512IFMA);
1077 if (HasLeaf7 && ((EBX >> 23) & 1))
1078 setFeature(X86::FEATURE_CLFLUSHOPT);
1079 if (HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save)
1080 setFeature(X86::FEATURE_AVX512PF);
1081 if (HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save)
1082 setFeature(X86::FEATURE_AVX512ER);
1083 if (HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save)
1084 setFeature(X86::FEATURE_AVX512CD);
1085 if (HasLeaf7 && ((EBX >> 29) & 1))
1086 setFeature(X86::FEATURE_SHA);
1087 if (HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save)
1088 setFeature(X86::FEATURE_AVX512BW);
1089 if (HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save)
1090 setFeature(X86::FEATURE_AVX512VL);
1091
1092 if (HasLeaf7 && ((ECX >> 1) & 1) && HasAVX512Save)
1093 setFeature(X86::FEATURE_AVX512VBMI);
1094 if (HasLeaf7 && ((ECX >> 6) & 1) && HasAVX512Save)
1095 setFeature(X86::FEATURE_AVX512VBMI2);
1096 if (HasLeaf7 && ((ECX >> 8) & 1))
1097 setFeature(X86::FEATURE_GFNI);
1098 if (HasLeaf7 && ((ECX >> 10) & 1) && HasAVX)
1099 setFeature(X86::FEATURE_VPCLMULQDQ);
1100 if (HasLeaf7 && ((ECX >> 11) & 1) && HasAVX512Save)
1101 setFeature(X86::FEATURE_AVX512VNNI);
1102 if (HasLeaf7 && ((ECX >> 12) & 1) && HasAVX512Save)
1103 setFeature(X86::FEATURE_AVX512BITALG);
1104 if (HasLeaf7 && ((ECX >> 14) & 1) && HasAVX512Save)
1105 setFeature(X86::FEATURE_AVX512VPOPCNTDQ);
1106
1107 if (HasLeaf7 && ((EDX >> 2) & 1) && HasAVX512Save)
1108 setFeature(X86::FEATURE_AVX5124VNNIW);
1109 if (HasLeaf7 && ((EDX >> 3) & 1) && HasAVX512Save)
1110 setFeature(X86::FEATURE_AVX5124FMAPS);
1111 if (HasLeaf7 && ((EDX >> 8) & 1) && HasAVX512Save)
1112 setFeature(X86::FEATURE_AVX512VP2INTERSECT);
1113
1114 bool HasLeaf7Subleaf1 =
1115 MaxLeaf >= 7 && !getX86CpuIDAndInfoEx(0x7, 0x1, &EAX, &EBX, &ECX, &EDX);
1116 if (HasLeaf7Subleaf1 && ((EAX >> 5) & 1) && HasAVX512Save)
1117 setFeature(X86::FEATURE_AVX512BF16);
1118
1119 unsigned MaxExtLevel;
1120 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX);
1121
1122 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 &&
1123 !getX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
1124 if (HasExtLeaf1 && ((ECX >> 6) & 1))
1125 setFeature(X86::FEATURE_SSE4_A);
1126 if (HasExtLeaf1 && ((ECX >> 11) & 1))
1127 setFeature(X86::FEATURE_XOP);
1128 if (HasExtLeaf1 && ((ECX >> 16) & 1))
1129 setFeature(X86::FEATURE_FMA4);
1130
1131 if (HasExtLeaf1 && ((EDX >> 29) & 1))
1132 setFeature(X86::FEATURE_64BIT);
1133 }
1134
getHostCPUName()1135 StringRef sys::getHostCPUName() {
1136 unsigned MaxLeaf = 0;
1137 const VendorSignatures Vendor = getVendorSignature(&MaxLeaf);
1138 if (Vendor == VendorSignatures::UNKNOWN)
1139 return "generic";
1140
1141 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
1142 getX86CpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
1143
1144 unsigned Family = 0, Model = 0;
1145 unsigned Features[(X86::CPU_FEATURE_MAX + 31) / 32] = {0};
1146 detectX86FamilyModel(EAX, &Family, &Model);
1147 getAvailableFeatures(ECX, EDX, MaxLeaf, Features);
1148
1149 // These aren't consumed in this file, but we try to keep some source code the
1150 // same or similar to compiler-rt.
1151 unsigned Type = 0;
1152 unsigned Subtype = 0;
1153
1154 StringRef CPU;
1155
1156 if (Vendor == VendorSignatures::GENUINE_INTEL) {
1157 CPU = getIntelProcessorTypeAndSubtype(Family, Model, Features, &Type,
1158 &Subtype);
1159 } else if (Vendor == VendorSignatures::AUTHENTIC_AMD) {
1160 CPU = getAMDProcessorTypeAndSubtype(Family, Model, Features, &Type,
1161 &Subtype);
1162 }
1163
1164 if (!CPU.empty())
1165 return CPU;
1166
1167 return "generic";
1168 }
1169
1170 #elif defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
getHostCPUName()1171 StringRef sys::getHostCPUName() {
1172 host_basic_info_data_t hostInfo;
1173 mach_msg_type_number_t infoCount;
1174
1175 infoCount = HOST_BASIC_INFO_COUNT;
1176 mach_port_t hostPort = mach_host_self();
1177 host_info(hostPort, HOST_BASIC_INFO, (host_info_t)&hostInfo,
1178 &infoCount);
1179 mach_port_deallocate(mach_task_self(), hostPort);
1180
1181 if (hostInfo.cpu_type != CPU_TYPE_POWERPC)
1182 return "generic";
1183
1184 switch (hostInfo.cpu_subtype) {
1185 case CPU_SUBTYPE_POWERPC_601:
1186 return "601";
1187 case CPU_SUBTYPE_POWERPC_602:
1188 return "602";
1189 case CPU_SUBTYPE_POWERPC_603:
1190 return "603";
1191 case CPU_SUBTYPE_POWERPC_603e:
1192 return "603e";
1193 case CPU_SUBTYPE_POWERPC_603ev:
1194 return "603ev";
1195 case CPU_SUBTYPE_POWERPC_604:
1196 return "604";
1197 case CPU_SUBTYPE_POWERPC_604e:
1198 return "604e";
1199 case CPU_SUBTYPE_POWERPC_620:
1200 return "620";
1201 case CPU_SUBTYPE_POWERPC_750:
1202 return "750";
1203 case CPU_SUBTYPE_POWERPC_7400:
1204 return "7400";
1205 case CPU_SUBTYPE_POWERPC_7450:
1206 return "7450";
1207 case CPU_SUBTYPE_POWERPC_970:
1208 return "970";
1209 default:;
1210 }
1211
1212 return "generic";
1213 }
1214 #elif defined(__linux__) && (defined(__ppc__) || defined(__powerpc__))
getHostCPUName()1215 StringRef sys::getHostCPUName() {
1216 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent();
1217 StringRef Content = P ? P->getBuffer() : "";
1218 return detail::getHostCPUNameForPowerPC(Content);
1219 }
1220 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
getHostCPUName()1221 StringRef sys::getHostCPUName() {
1222 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent();
1223 StringRef Content = P ? P->getBuffer() : "";
1224 return detail::getHostCPUNameForARM(Content);
1225 }
1226 #elif defined(__linux__) && defined(__s390x__)
getHostCPUName()1227 StringRef sys::getHostCPUName() {
1228 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent();
1229 StringRef Content = P ? P->getBuffer() : "";
1230 return detail::getHostCPUNameForS390x(Content);
1231 }
1232 #elif defined(__APPLE__) && defined(__aarch64__)
getHostCPUName()1233 StringRef sys::getHostCPUName() {
1234 return "cyclone";
1235 }
1236 #elif defined(__APPLE__) && defined(__arm__)
getHostCPUName()1237 StringRef sys::getHostCPUName() {
1238 host_basic_info_data_t hostInfo;
1239 mach_msg_type_number_t infoCount;
1240
1241 infoCount = HOST_BASIC_INFO_COUNT;
1242 mach_port_t hostPort = mach_host_self();
1243 host_info(hostPort, HOST_BASIC_INFO, (host_info_t)&hostInfo,
1244 &infoCount);
1245 mach_port_deallocate(mach_task_self(), hostPort);
1246
1247 if (hostInfo.cpu_type != CPU_TYPE_ARM) {
1248 assert(false && "CPUType not equal to ARM should not be possible on ARM");
1249 return "generic";
1250 }
1251 switch (hostInfo.cpu_subtype) {
1252 case CPU_SUBTYPE_ARM_V7S:
1253 return "swift";
1254 default:;
1255 }
1256
1257 return "generic";
1258 }
1259 #elif defined(_AIX)
getHostCPUName()1260 StringRef sys::getHostCPUName() {
1261 switch (_system_configuration.implementation) {
1262 case POWER_4:
1263 if (_system_configuration.version == PV_4_3)
1264 return "970";
1265 return "pwr4";
1266 case POWER_5:
1267 if (_system_configuration.version == PV_5)
1268 return "pwr5";
1269 return "pwr5x";
1270 case POWER_6:
1271 if (_system_configuration.version == PV_6_Compat)
1272 return "pwr6";
1273 return "pwr6x";
1274 case POWER_7:
1275 return "pwr7";
1276 case POWER_8:
1277 return "pwr8";
1278 case POWER_9:
1279 return "pwr9";
1280 // TODO: simplify this once the macro is available in all OS levels.
1281 #ifdef POWER_10
1282 case POWER_10:
1283 #else
1284 case 0x40000:
1285 #endif
1286 return "pwr10";
1287 default:
1288 return "generic";
1289 }
1290 }
1291 #else
getHostCPUName()1292 StringRef sys::getHostCPUName() { return "generic"; }
1293 namespace llvm {
1294 namespace sys {
1295 namespace detail {
1296 namespace x86 {
1297
getVendorSignature(unsigned * MaxLeaf)1298 VendorSignatures getVendorSignature(unsigned *MaxLeaf) {
1299 return VendorSignatures::UNKNOWN;
1300 }
1301
1302 } // namespace x86
1303 } // namespace detail
1304 } // namespace sys
1305 } // namespace llvm
1306 #endif
1307
1308 #if defined(__linux__) && (defined(__i386__) || defined(__x86_64__))
1309 // On Linux, the number of physical cores can be computed from /proc/cpuinfo,
1310 // using the number of unique physical/core id pairs. The following
1311 // implementation reads the /proc/cpuinfo format on an x86_64 system.
computeHostNumPhysicalCores()1312 int computeHostNumPhysicalCores() {
1313 // Enabled represents the number of physical id/core id pairs with at least
1314 // one processor id enabled by the CPU affinity mask.
1315 cpu_set_t Affinity, Enabled;
1316 if (sched_getaffinity(0, sizeof(Affinity), &Affinity) != 0)
1317 return -1;
1318 CPU_ZERO(&Enabled);
1319
1320 // Read /proc/cpuinfo as a stream (until EOF reached). It cannot be
1321 // mmapped because it appears to have 0 size.
1322 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> Text =
1323 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo");
1324 if (std::error_code EC = Text.getError()) {
1325 llvm::errs() << "Can't read "
1326 << "/proc/cpuinfo: " << EC.message() << "\n";
1327 return -1;
1328 }
1329 SmallVector<StringRef, 8> strs;
1330 (*Text)->getBuffer().split(strs, "\n", /*MaxSplit=*/-1,
1331 /*KeepEmpty=*/false);
1332 int CurProcessor = -1;
1333 int CurPhysicalId = -1;
1334 int CurSiblings = -1;
1335 int CurCoreId = -1;
1336 for (StringRef Line : strs) {
1337 std::pair<StringRef, StringRef> Data = Line.split(':');
1338 auto Name = Data.first.trim();
1339 auto Val = Data.second.trim();
1340 // These fields are available if the kernel is configured with CONFIG_SMP.
1341 if (Name == "processor")
1342 Val.getAsInteger(10, CurProcessor);
1343 else if (Name == "physical id")
1344 Val.getAsInteger(10, CurPhysicalId);
1345 else if (Name == "siblings")
1346 Val.getAsInteger(10, CurSiblings);
1347 else if (Name == "core id") {
1348 Val.getAsInteger(10, CurCoreId);
1349 // The processor id corresponds to an index into cpu_set_t.
1350 if (CPU_ISSET(CurProcessor, &Affinity))
1351 CPU_SET(CurPhysicalId * CurSiblings + CurCoreId, &Enabled);
1352 }
1353 }
1354 return CPU_COUNT(&Enabled);
1355 }
1356 #elif defined(__linux__) && defined(__powerpc__)
computeHostNumPhysicalCores()1357 int computeHostNumPhysicalCores() {
1358 cpu_set_t Affinity;
1359 if (sched_getaffinity(0, sizeof(Affinity), &Affinity) == 0)
1360 return CPU_COUNT(&Affinity);
1361
1362 // The call to sched_getaffinity() may have failed because the Affinity
1363 // mask is too small for the number of CPU's on the system (i.e. the
1364 // system has more than 1024 CPUs). Allocate a mask large enough for
1365 // twice as many CPUs.
1366 cpu_set_t *DynAffinity;
1367 DynAffinity = CPU_ALLOC(2048);
1368 if (sched_getaffinity(0, CPU_ALLOC_SIZE(2048), DynAffinity) == 0) {
1369 int NumCPUs = CPU_COUNT(DynAffinity);
1370 CPU_FREE(DynAffinity);
1371 return NumCPUs;
1372 }
1373 return -1;
1374 }
1375 #elif defined(__linux__) && defined(__s390x__)
computeHostNumPhysicalCores()1376 int computeHostNumPhysicalCores() { return sysconf(_SC_NPROCESSORS_ONLN); }
1377 #elif defined(__APPLE__) && defined(__x86_64__)
1378 #include <sys/param.h>
1379 #include <sys/sysctl.h>
1380
1381 // Gets the number of *physical cores* on the machine.
computeHostNumPhysicalCores()1382 int computeHostNumPhysicalCores() {
1383 uint32_t count;
1384 size_t len = sizeof(count);
1385 sysctlbyname("hw.physicalcpu", &count, &len, NULL, 0);
1386 if (count < 1) {
1387 int nm[2];
1388 nm[0] = CTL_HW;
1389 nm[1] = HW_AVAILCPU;
1390 sysctl(nm, 2, &count, &len, NULL, 0);
1391 if (count < 1)
1392 return -1;
1393 }
1394 return count;
1395 }
1396 #elif defined(__MVS__)
computeHostNumPhysicalCores()1397 int computeHostNumPhysicalCores() {
1398 enum {
1399 // Byte offset of the pointer to the Communications Vector Table (CVT) in
1400 // the Prefixed Save Area (PSA). The table entry is a 31-bit pointer and
1401 // will be zero-extended to uintptr_t.
1402 FLCCVT = 16,
1403 // Byte offset of the pointer to the Common System Data Area (CSD) in the
1404 // CVT. The table entry is a 31-bit pointer and will be zero-extended to
1405 // uintptr_t.
1406 CVTCSD = 660,
1407 // Byte offset to the number of live CPs in the LPAR, stored as a signed
1408 // 32-bit value in the table.
1409 CSD_NUMBER_ONLINE_STANDARD_CPS = 264,
1410 };
1411 char *PSA = 0;
1412 char *CVT = reinterpret_cast<char *>(
1413 static_cast<uintptr_t>(reinterpret_cast<unsigned int &>(PSA[FLCCVT])));
1414 char *CSD = reinterpret_cast<char *>(
1415 static_cast<uintptr_t>(reinterpret_cast<unsigned int &>(CVT[CVTCSD])));
1416 return reinterpret_cast<int &>(CSD[CSD_NUMBER_ONLINE_STANDARD_CPS]);
1417 }
1418 #elif defined(_WIN32) && LLVM_ENABLE_THREADS != 0
1419 // Defined in llvm/lib/Support/Windows/Threading.inc
1420 int computeHostNumPhysicalCores();
1421 #else
1422 // On other systems, return -1 to indicate unknown.
computeHostNumPhysicalCores()1423 static int computeHostNumPhysicalCores() { return -1; }
1424 #endif
1425
getHostNumPhysicalCores()1426 int sys::getHostNumPhysicalCores() {
1427 static int NumCores = computeHostNumPhysicalCores();
1428 return NumCores;
1429 }
1430
1431 #if defined(__i386__) || defined(_M_IX86) || \
1432 defined(__x86_64__) || defined(_M_X64)
getHostCPUFeatures(StringMap<bool> & Features)1433 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
1434 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
1435 unsigned MaxLevel;
1436
1437 if (getX86CpuIDAndInfo(0, &MaxLevel, &EBX, &ECX, &EDX) || MaxLevel < 1)
1438 return false;
1439
1440 getX86CpuIDAndInfo(1, &EAX, &EBX, &ECX, &EDX);
1441
1442 Features["cx8"] = (EDX >> 8) & 1;
1443 Features["cmov"] = (EDX >> 15) & 1;
1444 Features["mmx"] = (EDX >> 23) & 1;
1445 Features["fxsr"] = (EDX >> 24) & 1;
1446 Features["sse"] = (EDX >> 25) & 1;
1447 Features["sse2"] = (EDX >> 26) & 1;
1448
1449 Features["sse3"] = (ECX >> 0) & 1;
1450 Features["pclmul"] = (ECX >> 1) & 1;
1451 Features["ssse3"] = (ECX >> 9) & 1;
1452 Features["cx16"] = (ECX >> 13) & 1;
1453 Features["sse4.1"] = (ECX >> 19) & 1;
1454 Features["sse4.2"] = (ECX >> 20) & 1;
1455 Features["movbe"] = (ECX >> 22) & 1;
1456 Features["popcnt"] = (ECX >> 23) & 1;
1457 Features["aes"] = (ECX >> 25) & 1;
1458 Features["rdrnd"] = (ECX >> 30) & 1;
1459
1460 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
1461 // indicates that the AVX registers will be saved and restored on context
1462 // switch, then we have full AVX support.
1463 bool HasXSave = ((ECX >> 27) & 1) && !getX86XCR0(&EAX, &EDX);
1464 bool HasAVXSave = HasXSave && ((ECX >> 28) & 1) && ((EAX & 0x6) == 0x6);
1465 #if defined(__APPLE__)
1466 // Darwin lazily saves the AVX512 context on first use: trust that the OS will
1467 // save the AVX512 context if we use AVX512 instructions, even the bit is not
1468 // set right now.
1469 bool HasAVX512Save = true;
1470 #else
1471 // AVX512 requires additional context to be saved by the OS.
1472 bool HasAVX512Save = HasAVXSave && ((EAX & 0xe0) == 0xe0);
1473 #endif
1474 // AMX requires additional context to be saved by the OS.
1475 const unsigned AMXBits = (1 << 17) | (1 << 18);
1476 bool HasAMXSave = HasXSave && ((EAX & AMXBits) == AMXBits);
1477
1478 Features["avx"] = HasAVXSave;
1479 Features["fma"] = ((ECX >> 12) & 1) && HasAVXSave;
1480 // Only enable XSAVE if OS has enabled support for saving YMM state.
1481 Features["xsave"] = ((ECX >> 26) & 1) && HasAVXSave;
1482 Features["f16c"] = ((ECX >> 29) & 1) && HasAVXSave;
1483
1484 unsigned MaxExtLevel;
1485 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX);
1486
1487 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 &&
1488 !getX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
1489 Features["sahf"] = HasExtLeaf1 && ((ECX >> 0) & 1);
1490 Features["lzcnt"] = HasExtLeaf1 && ((ECX >> 5) & 1);
1491 Features["sse4a"] = HasExtLeaf1 && ((ECX >> 6) & 1);
1492 Features["prfchw"] = HasExtLeaf1 && ((ECX >> 8) & 1);
1493 Features["xop"] = HasExtLeaf1 && ((ECX >> 11) & 1) && HasAVXSave;
1494 Features["lwp"] = HasExtLeaf1 && ((ECX >> 15) & 1);
1495 Features["fma4"] = HasExtLeaf1 && ((ECX >> 16) & 1) && HasAVXSave;
1496 Features["tbm"] = HasExtLeaf1 && ((ECX >> 21) & 1);
1497 Features["mwaitx"] = HasExtLeaf1 && ((ECX >> 29) & 1);
1498
1499 Features["64bit"] = HasExtLeaf1 && ((EDX >> 29) & 1);
1500
1501 // Miscellaneous memory related features, detected by
1502 // using the 0x80000008 leaf of the CPUID instruction
1503 bool HasExtLeaf8 = MaxExtLevel >= 0x80000008 &&
1504 !getX86CpuIDAndInfo(0x80000008, &EAX, &EBX, &ECX, &EDX);
1505 Features["clzero"] = HasExtLeaf8 && ((EBX >> 0) & 1);
1506 Features["wbnoinvd"] = HasExtLeaf8 && ((EBX >> 9) & 1);
1507
1508 bool HasLeaf7 =
1509 MaxLevel >= 7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX);
1510
1511 Features["fsgsbase"] = HasLeaf7 && ((EBX >> 0) & 1);
1512 Features["sgx"] = HasLeaf7 && ((EBX >> 2) & 1);
1513 Features["bmi"] = HasLeaf7 && ((EBX >> 3) & 1);
1514 // AVX2 is only supported if we have the OS save support from AVX.
1515 Features["avx2"] = HasLeaf7 && ((EBX >> 5) & 1) && HasAVXSave;
1516 Features["bmi2"] = HasLeaf7 && ((EBX >> 8) & 1);
1517 Features["invpcid"] = HasLeaf7 && ((EBX >> 10) & 1);
1518 Features["rtm"] = HasLeaf7 && ((EBX >> 11) & 1);
1519 // AVX512 is only supported if the OS supports the context save for it.
1520 Features["avx512f"] = HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save;
1521 Features["avx512dq"] = HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save;
1522 Features["rdseed"] = HasLeaf7 && ((EBX >> 18) & 1);
1523 Features["adx"] = HasLeaf7 && ((EBX >> 19) & 1);
1524 Features["avx512ifma"] = HasLeaf7 && ((EBX >> 21) & 1) && HasAVX512Save;
1525 Features["clflushopt"] = HasLeaf7 && ((EBX >> 23) & 1);
1526 Features["clwb"] = HasLeaf7 && ((EBX >> 24) & 1);
1527 Features["avx512pf"] = HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save;
1528 Features["avx512er"] = HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save;
1529 Features["avx512cd"] = HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save;
1530 Features["sha"] = HasLeaf7 && ((EBX >> 29) & 1);
1531 Features["avx512bw"] = HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save;
1532 Features["avx512vl"] = HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save;
1533
1534 Features["prefetchwt1"] = HasLeaf7 && ((ECX >> 0) & 1);
1535 Features["avx512vbmi"] = HasLeaf7 && ((ECX >> 1) & 1) && HasAVX512Save;
1536 Features["pku"] = HasLeaf7 && ((ECX >> 4) & 1);
1537 Features["waitpkg"] = HasLeaf7 && ((ECX >> 5) & 1);
1538 Features["avx512vbmi2"] = HasLeaf7 && ((ECX >> 6) & 1) && HasAVX512Save;
1539 Features["shstk"] = HasLeaf7 && ((ECX >> 7) & 1);
1540 Features["gfni"] = HasLeaf7 && ((ECX >> 8) & 1);
1541 Features["vaes"] = HasLeaf7 && ((ECX >> 9) & 1) && HasAVXSave;
1542 Features["vpclmulqdq"] = HasLeaf7 && ((ECX >> 10) & 1) && HasAVXSave;
1543 Features["avx512vnni"] = HasLeaf7 && ((ECX >> 11) & 1) && HasAVX512Save;
1544 Features["avx512bitalg"] = HasLeaf7 && ((ECX >> 12) & 1) && HasAVX512Save;
1545 Features["avx512vpopcntdq"] = HasLeaf7 && ((ECX >> 14) & 1) && HasAVX512Save;
1546 Features["rdpid"] = HasLeaf7 && ((ECX >> 22) & 1);
1547 Features["kl"] = HasLeaf7 && ((ECX >> 23) & 1); // key locker
1548 Features["cldemote"] = HasLeaf7 && ((ECX >> 25) & 1);
1549 Features["movdiri"] = HasLeaf7 && ((ECX >> 27) & 1);
1550 Features["movdir64b"] = HasLeaf7 && ((ECX >> 28) & 1);
1551 Features["enqcmd"] = HasLeaf7 && ((ECX >> 29) & 1);
1552
1553 Features["uintr"] = HasLeaf7 && ((EDX >> 5) & 1);
1554 Features["avx512vp2intersect"] =
1555 HasLeaf7 && ((EDX >> 8) & 1) && HasAVX512Save;
1556 Features["serialize"] = HasLeaf7 && ((EDX >> 14) & 1);
1557 Features["tsxldtrk"] = HasLeaf7 && ((EDX >> 16) & 1);
1558 // There are two CPUID leafs which information associated with the pconfig
1559 // instruction:
1560 // EAX=0x7, ECX=0x0 indicates the availability of the instruction (via the 18th
1561 // bit of EDX), while the EAX=0x1b leaf returns information on the
1562 // availability of specific pconfig leafs.
1563 // The target feature here only refers to the the first of these two.
1564 // Users might need to check for the availability of specific pconfig
1565 // leaves using cpuid, since that information is ignored while
1566 // detecting features using the "-march=native" flag.
1567 // For more info, see X86 ISA docs.
1568 Features["pconfig"] = HasLeaf7 && ((EDX >> 18) & 1);
1569 Features["amx-bf16"] = HasLeaf7 && ((EDX >> 22) & 1) && HasAMXSave;
1570 Features["amx-tile"] = HasLeaf7 && ((EDX >> 24) & 1) && HasAMXSave;
1571 Features["amx-int8"] = HasLeaf7 && ((EDX >> 25) & 1) && HasAMXSave;
1572 bool HasLeaf7Subleaf1 =
1573 MaxLevel >= 7 && !getX86CpuIDAndInfoEx(0x7, 0x1, &EAX, &EBX, &ECX, &EDX);
1574 Features["avxvnni"] = HasLeaf7Subleaf1 && ((EAX >> 4) & 1) && HasAVXSave;
1575 Features["avx512bf16"] = HasLeaf7Subleaf1 && ((EAX >> 5) & 1) && HasAVX512Save;
1576 Features["hreset"] = HasLeaf7Subleaf1 && ((EAX >> 22) & 1);
1577
1578 bool HasLeafD = MaxLevel >= 0xd &&
1579 !getX86CpuIDAndInfoEx(0xd, 0x1, &EAX, &EBX, &ECX, &EDX);
1580
1581 // Only enable XSAVE if OS has enabled support for saving YMM state.
1582 Features["xsaveopt"] = HasLeafD && ((EAX >> 0) & 1) && HasAVXSave;
1583 Features["xsavec"] = HasLeafD && ((EAX >> 1) & 1) && HasAVXSave;
1584 Features["xsaves"] = HasLeafD && ((EAX >> 3) & 1) && HasAVXSave;
1585
1586 bool HasLeaf14 = MaxLevel >= 0x14 &&
1587 !getX86CpuIDAndInfoEx(0x14, 0x0, &EAX, &EBX, &ECX, &EDX);
1588
1589 Features["ptwrite"] = HasLeaf14 && ((EBX >> 4) & 1);
1590
1591 bool HasLeaf19 =
1592 MaxLevel >= 0x19 && !getX86CpuIDAndInfo(0x19, &EAX, &EBX, &ECX, &EDX);
1593 Features["widekl"] = HasLeaf7 && HasLeaf19 && ((EBX >> 2) & 1);
1594
1595 return true;
1596 }
1597 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
getHostCPUFeatures(StringMap<bool> & Features)1598 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
1599 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent();
1600 if (!P)
1601 return false;
1602
1603 SmallVector<StringRef, 32> Lines;
1604 P->getBuffer().split(Lines, "\n");
1605
1606 SmallVector<StringRef, 32> CPUFeatures;
1607
1608 // Look for the CPU features.
1609 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
1610 if (Lines[I].startswith("Features")) {
1611 Lines[I].split(CPUFeatures, ' ');
1612 break;
1613 }
1614
1615 #if defined(__aarch64__)
1616 // Keep track of which crypto features we have seen
1617 enum { CAP_AES = 0x1, CAP_PMULL = 0x2, CAP_SHA1 = 0x4, CAP_SHA2 = 0x8 };
1618 uint32_t crypto = 0;
1619 #endif
1620
1621 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) {
1622 StringRef LLVMFeatureStr = StringSwitch<StringRef>(CPUFeatures[I])
1623 #if defined(__aarch64__)
1624 .Case("asimd", "neon")
1625 .Case("fp", "fp-armv8")
1626 .Case("crc32", "crc")
1627 #else
1628 .Case("half", "fp16")
1629 .Case("neon", "neon")
1630 .Case("vfpv3", "vfp3")
1631 .Case("vfpv3d16", "d16")
1632 .Case("vfpv4", "vfp4")
1633 .Case("idiva", "hwdiv-arm")
1634 .Case("idivt", "hwdiv")
1635 #endif
1636 .Default("");
1637
1638 #if defined(__aarch64__)
1639 // We need to check crypto separately since we need all of the crypto
1640 // extensions to enable the subtarget feature
1641 if (CPUFeatures[I] == "aes")
1642 crypto |= CAP_AES;
1643 else if (CPUFeatures[I] == "pmull")
1644 crypto |= CAP_PMULL;
1645 else if (CPUFeatures[I] == "sha1")
1646 crypto |= CAP_SHA1;
1647 else if (CPUFeatures[I] == "sha2")
1648 crypto |= CAP_SHA2;
1649 #endif
1650
1651 if (LLVMFeatureStr != "")
1652 Features[LLVMFeatureStr] = true;
1653 }
1654
1655 #if defined(__aarch64__)
1656 // If we have all crypto bits we can add the feature
1657 if (crypto == (CAP_AES | CAP_PMULL | CAP_SHA1 | CAP_SHA2))
1658 Features["crypto"] = true;
1659 #endif
1660
1661 return true;
1662 }
1663 #elif defined(_WIN32) && (defined(__aarch64__) || defined(_M_ARM64))
getHostCPUFeatures(StringMap<bool> & Features)1664 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
1665 if (IsProcessorFeaturePresent(PF_ARM_NEON_INSTRUCTIONS_AVAILABLE))
1666 Features["neon"] = true;
1667 if (IsProcessorFeaturePresent(PF_ARM_V8_CRC32_INSTRUCTIONS_AVAILABLE))
1668 Features["crc"] = true;
1669 if (IsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE))
1670 Features["crypto"] = true;
1671
1672 return true;
1673 }
1674 #else
getHostCPUFeatures(StringMap<bool> & Features)1675 bool sys::getHostCPUFeatures(StringMap<bool> &Features) { return false; }
1676 #endif
1677
getProcessTriple()1678 std::string sys::getProcessTriple() {
1679 std::string TargetTripleString = updateTripleOSVersion(LLVM_HOST_TRIPLE);
1680 Triple PT(Triple::normalize(TargetTripleString));
1681
1682 if (sizeof(void *) == 8 && PT.isArch32Bit())
1683 PT = PT.get64BitArchVariant();
1684 if (sizeof(void *) == 4 && PT.isArch64Bit())
1685 PT = PT.get32BitArchVariant();
1686
1687 return PT.str();
1688 }
1689